Stan (software)

Stan is a probabilistic programming language that separates model specification from inference. Users write a model in the Stan language, which resembles C++ but includes blocks for data, parameters, transformed parameters, and model likelihood. The compiler translates this into C++ code, which is then compiled into an executable that performs Bayesian inference. The core inference engine implements Hamiltonian Monte Carlo, an efficient MCMC method that uses gradient information to explore the posterior distribution. The No-U-Turn Sampler (NUTS) automatically tunes step sizes and trajectory lengths, making it user-friendly for complex models. Stan also supports variational inference (ADVI) and penalized maximum likelihood estimation. Interfaces exist for R (rstan), Python (pystan, cmdstanpy), Julia, MATLAB, and Stata, among others.

Stan is used across many fields, including epidemiology, ecology, economics, psychology, and social sciences. It is particularly popular for hierarchical (multilevel) models, which are common in educational testing, clinical trials, and survey analysis. Stan has been employed in disease modeling, such as estimating COVID-19 transmission rates, and in ecological studies for species distribution modeling. In industry, companies use Stan for forecasting, customer analytics, and risk modeling. The Stan community maintains a rich ecosystem of resources, including the Stan User's Guide, a forum, and case studies. Stan is also integrated into higher-level tools like brms (Bayesian regression models using Stan) and rstanarm, which simplify model fitting for common regression scenarios.

Stan's key technical contributions include the efficient implementation of Hamiltonian Monte Carlo and the No-U-Turn Sampler, which automatically tunes HMC parameters without user intervention. The algorithm adapts step size and mass matrix during warmup, improving convergence and reducing manual tuning. Stan also implements a reverse-mode automatic differentiation engine, which computes gradients of the log-posterior with respect to parameters, enabling efficient sampling. The Stan Math Library is a standalone C++ library that provides automatic differentiation and matrix operations, used beyond the probabilistic programming context. Stan's design emphasizes numerical stability and performance, with support for sparse matrices and ODE solvers for dynamic models. These features make Stan a powerful tool for complex, high-dimensional Bayesian models.

Stan's name honors Stanislaw Ulam, a mathematician who co-developed the Monte Carlo method. The project began as a collaborative effort led by Andrew Gelman and others at Columbia University, with initial development by Bob Carpenter and colleagues. Stan supports a range of inference algorithms beyond HMC, including variational inference (ADVI) and Laplace approximation. It also includes a 'generated quantities' block for post-processing and predictive checks. Stan has a unique feature called 'reduce_sum' for parallelizing computations across cores, improving scalability. The Stan ecosystem includes tools like ShinyStan for interactive exploration of MCMC results. Stan's development is community-driven, with regular releases and a strong emphasis on reproducibility. Despite its power, Stan has a learning curve, and users often rely on wrapper packages to simplify model specification.

Hamiltonian Monte Carlo (HMC)

An MCMC method that uses gradient information to propose efficient moves in the parameter space.

No-U-Turn Sampler (NUTS)

An extension of HMC that automatically tunes step size and trajectory length, reducing manual tuning.

Automatic differentiation

A technique to compute derivatives of functions programmatically, used in Stan to obtain gradients of the log-posterior.

Variational inference

An approximate inference method that turns sampling into an optimization problem.